auto_label/llm/build_model.py

import torch
from langchain.llms.base import LLM
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig
from typing import Optional, Any, List
from langchain_core.callbacks import CallbackManagerForLLMRun
import os
from peft import PeftModel, PeftModelForCausalLM
from config.config import log
#自定义千问模型支持接口
class Qwen_LLM(LLM):
    tokenizer: AutoTokenizer = None
    model: AutoModelForCausalLM = None
    generation_config: GenerationConfig = None
    model_wrapper: PeftModelForCausalLM = None
    def __init__(self, model_path):
        super(Qwen_LLM, self).__init__()
        self.tokenizer = AutoTokenizer.from_pretrained(model_path)
        self.model = AutoModelForCausalLM.from_pretrained(model_path,
                                                          torch_dtype=torch.float16,
                                                          device_map={"":0})
        self.model_wrapper = PeftModel.from_pretrained(self.model, os.environ.get("MATH_LORA_PATH"), adapter_name=os.environ.get("MATH_LORA_NAME"))
        self.model_wrapper.load_adapter(os.environ.get("PHYSICS_LORA_PATH"), adapter_name=os.environ.get("PHYSICS_LORA_NAME"))
        self.generation_config = GenerationConfig(
            temperature=0.05,
            top_p=0.7,
            do_sample=True,
            max_new_tokens=600,  # max_length=max_new_tokens+input_sequence
            repetition_penalty=1.02,
            eos_token_id=self.tokenizer.eos_token_id
        )

    def _call(
        self,
        prompt: str,
        stop: Optional[List[str]] = None,
        run_manager: Optional[CallbackManagerForLLMRun] = None,
        **kwargs: Any,
    ) -> str:
        subject_id = kwargs.get("subject_id", -1)
        log.info("值："+str(subject_id))
        #物理学科
        if subject_id == 12:
            #self.model_wrapper.set_adapter(os.environ.get("PHYSICS_LORA_NAME"))
            PeftModel.from_pretrained(self.model, os.environ.get("PHYSICS_LORA_PATH"),
                                      adapter_name=os.environ.get("PHYSICS_LORA_NAME"))
        if subject_id == 3:
            self.model_wrapper = PeftModel.from_pretrained(self.model, os.environ.get("MATH_LORA_PATH"), adapter_name=os.environ.get("MATH_LORA_NAME"))
        with torch.no_grad():
            ids = self.tokenizer.encode(prompt) + [self.tokenizer.eos_token_id]
            input_ids = torch.tensor([ids]).cuda()
            output = self.model_wrapper.generate(input_ids=input_ids,
                                         generation_config=self.generation_config)
            out_ids = output.cpu()[0][input_ids.size(1):]
        answer = self.tokenizer.decode(out_ids, skip_special_tokens=True)
        log.info(answer)
        return answer

    def _llm_type(self) -> str:
        """Return type of llm"""
        return "qwen2"

qwen2 = Qwen_LLM(os.environ.get("MODEL_PATH"))